The Price of Anarchy in Cooperative Network Creation Games

13 Pages·2017·0.55 MB·English

The Price of Anarchy in Cooperative Network Creation
Games
Erik D. Demaine, Mohammadtaghi Hajiaghayi, Hamid Mahini, Morteza
Zadimoghaddam
To cite this version:
Erik D. Demaine, Mohammadtaghi Hajiaghayi, Hamid Mahini, Morteza Zadimoghaddam. The Price
of Anarchy in Cooperative Network Creation Games. 26th International Symposium on Theoretical
AspectsofComputerScienceSTACS2009,Feb2009,Freiburg,Germany. pp.301-312. ￿inria-00359313￿
HAL Id: inria-00359313
https://hal.inria.fr/inria-00359313
Submitted on 9 Feb 2009
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est
archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
SymposiumonTheoreticalAspectsofComputerScience2009(Freiburg),pp.301–312
www.stacs-conf.org
THE PRICE OF ANARCHY IN
COOPERATIVE NETWORK CREATION GAMES
ERIK D.DEMAINE1 ANDMOHAMMADTAGHIHAJIAGHAYI1,2 ANDHAMIDMAHINI3,4 AND
MORTEZA ZADIMOGHADDAM4
1 MIT Computer Science and Artificial Intelligence Laboratory, 32 Vassar St., Cambridge, MA
02139, USA
E-mail address: [email protected]
2 AT&TLabs — Research, 180 Park Ave., Florham Park, NJ 07932, USA
E-mail address: [email protected]
3 School of Computer Science, Institutefor Theoretical Physics and Mathematics, Tehran
E-mail address: [email protected]
4 Department of Computer Engineering, Sharif University of Technology
E-mail address: [email protected]
Abstract. We analyze thestructure of equilibria and the price of anarchy in the family
of network creation games considered extensively in thepast few years, which attempt to
unify the network design and network routing problems by modeling both creation and
usagecosts. Ingeneral,thegamesareplayedonahostgraph,whereeachnodeisaselfish
independent agent (player) and each edge has a fixed link creation cost α. Together the
agentscreateanetwork(asubgraphofthehost graph)whileselfishly minimizingthelink
creationcostsplusthesumofthedistancestoallotherplayers(usagecost). Inthispaper,
we pursuetwo important facets of the network creation game.
First,westudyextensivelyanaturalversionofthegame,calledthecooperativemodel,
where nodes can collaborate and share the cost of creating any edge in the host graph.
We prove the first nontrivial bounds in this model, establishing that the price of anarchy
ispolylogarithmic innfor all valuesof αin completehost graphs. This boundisthefirst
result of this type for any version of the network creation game; most previous general
upper bounds are polynomial in n. Interestingly, we also show that equilibrium graphs
havepolylogarithmic diameter for themost natural range of α (at most npolylgn).
Second,westudytheimpactofthenaturalassumptionthatthehostgraphisageneral
graph, not necessarily complete. This model is a simple example of nonuniform creation
costs among the edges (effectively allowing weights of α and ∞). We prove the first
assemblageofupperandlowerboundsforthiscontext,establishingnontrivialtightbounds
formanyrangesofα,forboththeunilateralandcooperativeversionsofnetworkcreation.
In particular, we establish polynomial lower bounds for both versions and many ranges
ofα,evenforthissimplenonuniformcost model, whichsharplycontraststheconjectured
constant boundsfor these games in complete (uniform) graphs.
(cid:13)c E.D.Demaine,M.Hajiaghayi,H.Mahini,andM.Zadimoghaddam
(cid:13)CC CreativeCommonsAttribution-NoDerivsLicense
302 E.D.DEMAINE,M.HAJIAGHAYI,H.MAHINI,ANDM.ZADIMOGHADDAM
1. Introduction
A fundamental family of problems at the intersection between computer science and
operations research is network design. This area of research has become increasingly impor-
tant given the continued growth of computer networks such as the Internet. Traditionally,
we want to find a minimum-cost (sub)network that satisfies some specified property such
as k-connectivity or connectivity on terminals (as in the classic Steiner tree problem). This
goal captures the (possibly incremental) creation cost of the network, but does not incor-
porate the cost of actually using the network. In contrast, network routing has the goal of
optimizing the usage cost of the network, but assumes that the network has already been
created.
Network creation games attempt to unify the network design and network routing
problems by modeling both creation and usage costs. In general, the game is played on a
host graph, where each node is an independent agent (player), and the goal is to create a
network from a subgraph of the host graph. Collectively, the nodes decide which edges of
the host graph are worth creating as links in the network. Every link has the same creation
cost α. (Equivalently, links have creation costs of α and , depending on whether they are
∞
edges of the host graph.) In addition to these creation costs, each node incurs a usage cost
equal to the sum of distances to all other nodes in the network. Equivalently, if we divide
the cost (and thus α) by the number n of nodes, the usage cost for each node is its average
distance to all other nodes. (This natural cost model has been used in, e.g., contribution
games and network-formation games.)
There are several versions of the network creation game that vary how links are pur-
chased. Intheunilateral model—introducedby Fabrikant, Luthra,Maneva, Papadimitriou,
and Shenker [15]—every node (player) can locally decide to purchase any edge incident to
the node in the host graph, at a cost of α. In the bilateral model—introduced by Corbo
and Parkes [9]—both endpoints of an edge must agree beforethey can create a link between
them, and the two nodes share the α creation cost equally. In the cooperative model—
introduced by Albers, Eilts, Even-Dar, Mansour, and Roditty [2]—any node can purchase
any amount of any edge in the host graph, and a link gets created when the total purchased
amount is at least α.
Tomodelthedominantbehavioroflarge-scalenetworkingscenariossuchastheInternet,
we consider the case where every node (player) selfishly tries to minimize its own creation
and usage cost [18, 15, 2, 9]. This game-theoretic setting naturally leads to the various
kinds of equilibria and the study of their structure. Two frequently considered notions are
Nash equilibrium [24, 25], where no player can change its strategy (which edges to buy)
to locally improve its cost, and strong Nash equilibrium [6, 3, 1], where no coalition of
players can change their collective strategy to locally improve the cost of each player in
the coalition. Nash equilibria capture the combined effect of both selfishness and lack of
coordination, while strong Nash equilibria separates these issues, enablingcoordination and
capturing the specific effect of selfishness. However, the notion of strong Nash equilibrium
is extremely restrictive in our context, because all players can simultaneously change their
entire strategies, abusing the local optimality intended by original Nash equilibria, and
effectively forcing globally near-optimal solutions [3].
We consider weaker notions of equilibria, which broadens the scope of equilibria and
therefore strengthens our upper bounds, where players can change their strategy on only a
singleedgeatatime. Inacollaborative equilibrium, even coalitions ofplayers donotwishto
THE PRICE OF ANARCHY IN COOPERATIVE NETWORK CREATION GAMES 303
changetheircollective strategy onany singleedge; thisconceptisparticularly importantfor
the cooperative network creation game, wheremultiple players must negotiate their relative
valuations of an edge. (This notion is the natural generalization of pairwise stability from
[9]to arbitrary cost sharing.) Collaborative equilibriaare essentially acompromise between
Nash and strong Nash equilibria: they still enable coordination among players and thus
capture the specific effect of selfishness, like strong Nash, yet they consider more local
moves, in the spirit of Nash. In particular, any results about all collaborative equilibria
also apply to all strong Nash equilibria. Collaborative equilibria also make more sense
computationally: players canefficiently detectequilibriumusingasimplebiddingprocedure
(whereas this problem is NP-hard for strong Nash), and the resulting dynamics converge to
such equilibria (see Section 2.2).
The structure of equilibria in network creation games is not very well understood. For
example, Fabrikant et al. [15] conjectured that equilibrium graphs in the unilateral model
were all trees, but this conjecture was disproved by Albers et al. [2]. One particularly
interesting structural feature is whether all equilibrium graphs have small diameter (say,
polylogarithmic), analogous to the small-world phenomenon [19, 14], In the original uni-
lateral version of the problem, the best general lower bound is just a constant and the
best general upper bound is polynomial. A closely related issue is the price of anarchy
[20, 26, 28], that is, the worst possible ratio of the total cost of an equilibrium (found by
independent selfish behavior) and the optimal total cost possible by a centralized solution
(maximizing social welfare). The price of anarchy is a well-studied concept in algorithmic
game theory for problems such as load balancing, routing, and network design; see, e.g.,
[26, 10, 27, 15, 5, 4, 8, 9, 2, 11]. Upper bounds on diameter of equilibrium graphs translate
toapproximately equalupperboundsonthepriceofanarchy, butnotnecessarily viceversa.
In the unilateral version, for example, there is a general 2O(√lgn) upper bound on the price
of anarchy.
Previous work. Network creation games have been studied extensively in the literature
since their introduction in 2003.
Fortheunilateralversionandacompletehostgraph,Fabrikantetal.[15]proveanupper
bound of O(√α) on the price of anarchy for all α. Lin [23] proves that the price of anarchy
is constant for two ranges of α: α = O(√n) and α cn3/2 for some c > 0. Independently,
≥
Albers et al. [2] prove that the price of anarchy is constant for α = O(√n), as well as for
the larger range α 12n lgn . In addition, Albers et al. prove a general upper bound of
≥ ⌈ ⌉
15 1+(min α2,n2 )1/3 . The latter bound shows the first sublinear worst-case bound,
(cid:16) { n α } (cid:17)
O(n1/3), for all α. Demaine et al. [11] prove the first o(nε) upper bound for general α,
namely, 2O(√lgn). They also prove a constant upper bound for α = O(n1 ε) for any fixed
−
ε > 0, and improvetheconstant upperboundby Alberset al. (withthelead constant of 15)
to 6 for α < (n/2)1/2 and to 4 for α < (n/2)1/3. Andelmen et al. [3] show that, among
strong Nash equilibria, the price of anarchy is at most 2.
For the bilateral version and a complete host graph, Corbo and Parkes [9] prove that
the price of anarchy is between Ω(lgα) and O(min √α,n/√α). Demaine et al. [11] prove
{
that the upper bound is tight, establishing the price of anarchy to be Θ(min √α,n/√α )
{ }
in this case.
For the cooperative version and a complete host graph, the only known result is an
upper bound of 15 1+(min α2, n2 )1/3 , proved by Albers et al. [2].
(cid:16) { n α } (cid:17)
304 E.D.DEMAINE,M.HAJIAGHAYI,H.MAHINI,ANDM.ZADIMOGHADDAM
Other variations of network creation games allow nonuniform interests in connectivity
between nodes [17] and nodes with limited budgets for buying edges [21].
Our results. Our research pursues two important facets of the network creation game.
First, we make an extensive study of a natural version of the game—the cooperative
model—where the only previous results were simple extensions from unilateral analysis.
We substantially improve the bounds in this case, showing that the price of anarchy is
polylogarithmic in n for all values of α in complete graphs. This is the first result of this
type for any version of the network creation game. As mentioned above, this result applies
to both collaborative equilibria and strong Nash equilibria. Interestingly, we also show that
equilibrium graphs have polylogarithmic diameter for the most natural range of α (at most
npolylgn). Note that, because of the locally greedy nature of Nash equilibria, we cannot
usetheclassic probabilisticspanning(sub)treeembeddingmachinery of [7,16,13]toobtain
polylogarithmic bounds (although this machinery can be applied to approximate the global
social optimum).
Second, we study the impact of the natural assumption that the host graph is a general
graph, not necessarily complete, inspired by practical limitations in constructing network
links. This model is a simple example of nonuniform creation costs among the edges (ef-
fectively allowing weights of α and ). Surprisingly, no bounds on the diameter or the
∞
price of anarchy have been proved before in this context. We prove several upper and lower
bounds, establishing nontrivial tight bounds for many ranges of α, for both the unilateral
and cooperative versions. In particular, we establish polynomial lower bounds for both
versions and many ranges of α, even for this simple nonuniform cost model. These results
are particularly interesting because, by contrast, no superconstant lower bound has been
shown for either game in complete (uniform) graphs. Thus, while we believe that the price
of anarchy is polylogarithmic (or even constant) for complete graphs, we show a significant
departure from this behavior in general graphs.
Our proof techniques are most closely related in spirit to “region growing” from ap-
proximation algorithms; see, e.g., [22]. Our general goal is to prove an upper bound on
diameter by way of an upper bound on the expansion of the graph. However, we have not
beenable to get suchan argument to work directly in general. Themain difficulty is that, if
we imagine buildingabreadth-first-search tree from anode, then connecting that root node
to another node does not necessarily benefit the node much: it may only get closer to a
small fraction of nodes in the BFS subtree. Thus, no node is motivated selfishly to improve
the network, so several nodes must coordinate their changes to make improvements. The
cooperative version of the game gives us some leverage to address this difficulty. We hope
that this approach, particularly the structure we prove of equilibria, will shed some light on
the still-open unilateral version of the game, where the best bounds on the price of anarchy
are Ω(1) and 2O(√lgn).
Table 1 summarizes our results. Section 4 proves our polylogarithmic upper bounds
on the price of anarchy for all ranges of α in the cooperative network creation game in
complete graphs. Section 5 considers how the cooperative network creation game differs in
general graphs, and proves ourupperboundsfor this model. Section 6extends these results
to apply to the unilateral network creation game in general graphs. Section 7 proves lower
bounds for both the unilateral and cooperative network creation games in general graphs,
which match our upper bounds for some ranges of α.
THE PRICE OF ANARCHY IN COOPERATIVE NETWORK CREATION GAMES 305
α= 0 n nlg0.52n nlg7.16n n3/2 n5/3 n2 n2lgn ∞
Cooperative, complete graph Θ(1) lg3.32n O`lgn+pnlg3.58n´ Θ(1)
Cooperative, general graph O(α1/3) O(n1/3)α,Ω(pα) Θ(n2) O`n2 lgn´ Θ(1)
Unilateral, general graph O(α1/2) O(n1/2), Ω(α) n Θ(n2)α α Θ(1)
n α
Table 1: Summary of our bounds on equilibrium diameter and price of anarchy for cooper-
ative network creation in complete graphs,and unilateral andcooperative network
creation in general graphs. For all three of these models, our bounds are strict
improvements over the best previous bounds.
2. Models
In this section, we formally define the different models of the network creation game.
2.1. Unilateral Model
We start with the unilateral model, introduced in [15]. The game is played on a host
graph G = (V,E). Assume V = 1,2,...,n . We have n players, one per vertex. The
{ }
strategy ofplayer iisspecifiedbyasubsets of j : i,j E , definingthesetofneighbors
i
{ { } ∈ }
to which player i creates a link. Thus each player can only create links corresponding to
edges incident to node i in the host graph G Together, let s = s ,s ,...,s denote the
1 2 n
h i
joint strategy of all players.
To define the cost of strategies, we introduce a spanning subgraph G of the host
s
graph G. Namely, G has an edge i,j E(G) if either i s or j s . Define d (i,j) to
s { } ∈ ∈ j ∈ i Gs
be the distance between vertices i and j in graph G . Then the cost incurred by player i is
s
c (s)= α s + n d (i,j).Thetotalcostincurredbyjointstrategysisc(s)= n c (s).
i | i| Pj=1 Gs Pi=1 i
A (pure) Nash equilibrium is a joint strategy s such that c (s) c (s) for all joint
i i ′
≤
strategies s that differ from s in only one player i. The price of anarchy is then the
′
maximum cost of a Nash equilibrium divided by the minimum cost of any joint strategy
(called the social optimum).
2.2. Cooperative Model
Next we turn to the cooperative model, introduced in [15, 2]. Again, the game is
played on a host graph G = (V,E), with one player per vertex. Assume V = 1,2,...,n
{ }
and E = e ,e ,...,e . Now the strategy of player i is specified by a vector s =
1 2 E i
{ | |}
s(i,e ),s(i,e ),...,s(i,e ) , wheres(i,e )correspondstothevaluethatplayer iiswilling
1 2 E j
h | | i
to pay for link e . Together, s= s ,s ,...,s denotes the strategies of all players.
j 1 2 n
h i
We define a spanning subgraph G = (V,E ) of the host graph G: e is an edge
s s j
of G if s(i,e ) α. To make the total cost for an edge e exactly 0 or α in
s Pi∈V(G) j ≥ j
all cases, if s(i,e ) > α, we uniformly scale the costs to sum to α: s(i,e ) =
Pi∈V(G) j ′ j
αs(i,e )/ s(k,e ) (Equilibria will always have s = s.) Then the cost incurred by
playerji isPc (ks∈)V=(G) j s(i,e )+ n d (i,j). The total c′ost incurredby joint strategy
s is c(s) = αi E +Pejn∈Es ′n jd (Pi,jj)=.1 Gs
| s| Pi=1Pj=1 Gs
In this cooperative model, the notion of Nash equilibrium is less natural because it
allowsonlyoneplayertochangestrategy, whereasacooperativepurchaseingeneralrequires
many players to change their strategy. Therefore we use a stronger notion of equilibrium
that allows coalition among players, inspired by the strong Nash equilibrium of Aumann
306 E.D.DEMAINE,M.HAJIAGHAYI,H.MAHINI,ANDM.ZADIMOGHADDAM
[6], and modeled after the pairwise stability property introduced for the bilateral game
[9]. Namely, a joint strategy s is collaboratively equilibrium if, for any edge e of the host
graph G, for any coalition C V, for any joint strategy s differing from s in only s(i,e)
′ ′
⊆
for i C, some player i C has c (s) > c (s). Note that any such joint strategy must have
i ′ i
∈ ∈
every sum s(i,e ) equal to either 0 or α, so we can measure the cost c (s) in terms
Pi∈V(G) j i
of s(i,e ) instead of s(i,e ). The price of anarchy is the maximum cost of a collaborative
j ′ j
equilibrium divided by the minimum cost of any joint strategy (the social optimum).
We can define a simple dynamics for the cooperative network creation game in which
we repeatedly pick a pair of vertices, have all players determine their valuation of an edge
between those vertices (change in c (s) from addition or removal), and players thereby bid
i
on the edge and change their strategies. These dynamics always converge to a collaborative
equilibrium because each change decreases the total cost c(s), which is a discrete quantity
in the lattice Z + αZ. Indeed, the system therefore converges after a number of steps
polynomial in n and the smallest integer multiple of α (if one exists). More generally,
we can show an exponential upper bound in terms of just n by observing that the graph
uniquely determines c(s), so we can never repeat a graph by decreasing c(s).
3. Preliminaries
In this section, we define some helpful notation and prove some basic results. Call a
graph G corresponding to an equilibrium joint strategy s an equilibrium graph. In such
s
a graph, let d (u,v) be the length of the shortest path from u to v and Dist (u) be
Gs Gs
d (u,v). Let N (u) denote the set of vertices with distance at most k from
Pv∈V(Gs) Gs k
vertex u, and let N = min N (v). In both the unilateral and cooperative network
k v G k
∈ | |
creation games, the total cost of a strategy consists of two parts. We refer to the cost of
buying edges as the creation cost and the cost d (u,v) as the usage cost.
FirstweprovetheexistenceofcollaborativePeqvu∈iVli(bGrsi)afGorscompletehostgraphs. Similar
results are known in the unilateral case [15, 3].
Lemma 3.1. In the cooperative network creation game, any complete graph is a collabora-
tive equilibrium for α 2, and any star graph is a collaborative equilibrium for α 2.
≤ ≥
Next we show that, in the unilateral version, a bound on the usage cost suffices to
bound the total cost of an equilibrium graph G , similar to [11, Lemma 1].
s
Lemma 3.2. The total cost of any equilibrium graph in the unilateral game is at most
αn+2 d (u,v).
Pu,v∈V(Gs) Gs
Next we prove a more specific bound for the cooperative version, using the following
bound on the number of edges in a graph of large girth:
Lemma 3.3. [12] The number of edges in an n-vertex graph of odd girth g is O(n1+2/(g 1)).
−
Lemma 3.4. For any integer g, the total cost of any equilibrium graph G is at most
s
αO(n1+2/g)+g d (u,v).
Pu,v∈V(Gs) Gs
THE PRICE OF ANARCHY IN COOPERATIVE NETWORK CREATION GAMES 307
4. Cooperative Version in Complete Graphs
Inthissection, westudythepriceofanarchy whenanynumberofplayers cancooperate
to create any link, and the host graph is the complete graph.
We start with two lemmata that hold for both the unilateral and cooperative versions
of the problem. The firstlemma boundsa kind of “doublingradius” of large neighborhoods
around any vertex, which the second lemma uses to bound the usage cost.
Lemma 4.1. [11, Lemma4] For any vertex u in an equilibrium graph G , if N (u) > n/2,
s k
| |
then N (u) n.
2k+2α/n
| | ≥
Lemma 4.2. If we have N (u) > n/2 for some vertex u in an equilibrium graph G , the
k s
usage cost is at most O(n2k+αn).
Next we show how to improve the bound on “doubling radius” for large neighborhoods
in the cooperative game:
Lemma 4.3. For any vertex u in an equilibrium graph G , if N (u) > n/2, then
s k
| |
N (u) n.
| 2k+4√α/n | ≥
Next we consider what happens with arbitrary neighborhoods, using techniques similar
to [11, Lemma 5].
Lemma 4.4. If N (u) Y for every vertex u in an equilibrium graph G , then either
k s
| | ≥
N (u) > n/2 for some vertex u or N (u) Y2n/α for every vertex u.
4k+2 5k+3
| | | | ≥
Proof. If there is a vertex u with N (u) > n/2, then the claim is obvious. Otherwise,
4k+2
| |
for every vertex u, N (u) n/2. Let u be an arbitrary vertex. Let S be the set of
4k+2
| | ≤
vertices whose distance from u is 4k+3. We select a subset of S, called center points, by
the following greedy algorithm. We repeatedly select an unmarked vertex z S as a center
∈
point, mark all unmarked vertices in S whose distance from z is at most 2k, and assign
these vertices to z.
Suppose that we select l vertices
x ,x ,...,x as center points. We prove S
1 2 l
that l N (u)n/α. Let C be the ver-
k i
≥ | |
tices in S assigned to x ; see Figure 1. By
i
construction, S = l C . We also assign y
Si=1 i
eachvertexv atdistanceatleast4k+4from w
u to one of these center points, as follows. x
Pick any one shortest path from v to u that
x
i
contains some vertex w S, and assign v u
∈ C
to the same center point as w. This vertex i
w is unique in this path because this path
is a shortest path from v to u. Let T be
i x
k
the set of vertices assigned to x and whose
i
distance from u is more than 4k + 2. By Ck
construction, l T is the set of vertices
Si=1 i
at distance more than 4k +2 from u. The
shortest path from v T to u uses some Figure 1: Center points.
i
∈
vertex w C . For any vertex x whose dis-
i
∈
tance is at most k from u and for any y T , adding the edge u,x decreases the distance
i i
∈ { }
308 E.D.DEMAINE,M.HAJIAGHAYI,H.MAHINI,ANDM.ZADIMOGHADDAM
between x and y at least 2, because the shortest path from y T to u uses some vertex
i
∈
w C , as shown in Figure 1. By adding edge u,x , the distance between u and w would
i i
∈ { }
become at most 2k +1 and the distance between x and w would become at most 3k +1,
where x is any vertex whose distance from u is at most k. Because the current distance be-
tween xand w is at least 4k+3 k = 3k+3, addingthe edge u,x decreases this distance
i
− { }
by at least 2. Consequently the distance between x and any y T decreases by at least 2.
i
∈
Note that the distance between x and y is at least d (u,y) k, and after adding edge
Gs −
(u,x ), this distance becomes at most 3k+1+d (w,y) = 3k+1+d (u,y) d (u,w) =
i Gs Gs − Gs
3k+1+d (u,y) (4k+3) = d (u,y) k 2.
Gs − Gs − −
Thus any vertex y T has incentive to pay at least 2 N (u) for edge u,x . Because
i k i
∈ | | { }
theedge u,x isnotinequilibrium,weconcludethatα 2T N (u). Ontheotherhand,
i i k
{ } ≥ | || |
N (u) n/2, so l T n/2. Therefore, lα 2N (u) l T n N (u) and
| 4k+2 | ≤ Pi=1| i| ≥ ≥ | k |Pi=1| i| ≥ | k |
hence l n N (u)/α.
k
≥ | |
According to the greedy algorithm, the distance between any pair of center points is
more than 2k; hence, N (x ) N (x ) = for i = j. By the hypothesis of the lemma,
k i k j
∩ ∅ 6
N (x ) Y for every vertex x ; hence l N (x ) = l N (x ) lY. For every
| k i | ≥ i |Si=1 k i | Pi=1| k i | ≥
i l, we have d (u,x ) = 4k+3, so vertex u has a path of length at most 5k+3 to every
≤ Gs i
vertex whose distance to x is at most k. Therefore, N (u) l N (x ) lY
Yn N (u)/α Y2n/α. i | 5k+3 | ≥ |Si=1 k i | ≥ ≥2
k
| | ≥
Now we are ready to prove bounds on the price of anarchy. We start with the case
when α is a bit smaller than n:
Theorem 4.5. For 1 α < n1 ε, the price of anarchy is at most O(1/ε1+lg5).
−
≤
Next we prove a polylogarithmic bound on the price of anarchy when α is close to n.
Theorem 4.6. For α = O(n), the price of anarchy is O(lg1+lg5n) and the diameter of any
equilibrium graph is O(lglg5n).
Proof. Consider an equilibrium graph G . The proof is similar to the proof of Theorem 4.5.
s
Define a = max 2,2α/n +1 and a = 5a +3, or equivalently a = 4a1+3 5i 3 <
a 5i, for1all i >{1. By }Lemma 4.4,i for eai−ch1 i 1, either N (iv) >2n0/2· for−so4me
1 ≥ 4ai+2
vertex v or N (n/α)N2. Let j be the least number for which N (v) > n/2
ai+1 ≥ ai | 4aj+2 |
for some vertex v. By this definition, for each i < j, N (n/α)N2. Because N >
ai+1 ≥ ai a1
2max 1,α/n , we obtain that N > 22i−1max 1,α/n for every i j. On the other
{ } ai { } ≤
hand, 22j−1 22j−1max 1,α/n < N n, so j < lglgn + 1 and a < a 5lglgn+1 <
≤ { } aj ≤ j 1
(2 + 2α/n + 1 + 1)5lglg5n = 10(2 + α/n)lglg5n. Therefore N (v) >
4[10(2+α/n)lglg5n]+2
·
n/2 for some vertex v and using Lemma 4.1, we conclude that the distance of v to all
other vertices is at most 2[40(2 + α/n)lglg5n + 2] + 2α/n. Thus the diameter of G is
s
at most O((1 + α/n)lglg5n). Setting g = lgn in Lemma 3.4, the cost of G is at most
s
αO(n) + (lgn)O(n2(1 + α/n)lglg5n) = O((αn + n2)lg1+lg5n). Therefore the price of
anarchy is at most O(lg1+lg5n). 2
When α is a bit larger than n, we can obtain a constant bound on the price of anarchy.
First we need a somewhat stronger result on the behavior of neighborhoods:
Lemma 4.7. If N (u) Y for every vertex u in an equilibrium graph G , then either
k s
| | ≥
N (u) > n/2 for some vertex u or N (u) Y2kn/2α for every vertex u.
5k 6k+1
| | | | ≥
THE PRICE OF ANARCHY IN COOPERATIVE NETWORK CREATION GAMES 309
Theorem 4.8. For any α > n, the price of anarchy is O( n/αlg1+lg6n) and the diameter
p
of any equilibrium graph is O(lglg6n α/n).
·p
By Theorem 4.8, we conclude the following:
Corollary 4.9. For α = Ω(nlg2+2lg6n) Ω(nlg7.16n), the price of anarchy is O(1).
≈
5. Cooperative Version in General Graphs
In this section, we study the price of anarchy when only some links can be created,
e.g., because of physical limitations. In this case, the social optimum is no longer simply a
clique or a star.
We start by bounding the growth of distances from the host graph G to an arbitrary
equilibrium graph G :
s
Lemma 5.1. For any two vertices u and v in any equilibrium graph G , d (u,v) =
s Gs
O(d (u,v)+α1/3d (u,v)2/3).
G G
Proof. Let u= v ,v ,...,v = v be a shortest path in G between u and v, so k = d (u,v).
0 1 k G
Suppose that the distance between v and v in G is d , for 0 i k. We first prove
0 i s i
≤ ≤
that d d + 1 + 9α/d for 0 i < k. If edge v ,v already exists in G , the
i+1 i i i i+1 s
≤ p ≤ { }
inequality clearly holds. Otherwise, adding this edge decreases the distance between x and
y by at least di+13−di, where x is a vertex whose distance is at most di+13−di −1 from vi+1
and y is a vertex in a shortest path from v to v . Therefore any vertex x whose distance is
i 0
at most di+13−di −1 from vi+1 can pay di+13−didi for this edge. Because this edge does not
exist in Gs and because there are at least di+13−di vertices of distance at most di+13−di −1
2
from vi+1, we conclude that (cid:16)di+13−di(cid:17) di ≤ α. Thus we have di+1 ≤ di +1+p9α/di for
0 i < k. Next we prove that d d + 1 + 5α1/3. If edge v ,v already exists
i+1 i i i+1
≤ ≤ { }
in G , the inequality clearly holds. Otherwise, adding this edge decreases the distance
s
between z and w by at least di+1−di, where z and w are two vertices whose distances from
5
2
vi+1 and vi, respectively, are less than di+15−di. There are at least at least (cid:16)di+15−di(cid:17) pair
of vertices like (z,w). Because the edge v ,v does not exist in G , we conclude that
i i+1 s
{ }
3
(cid:16)di+15−di(cid:17) ≤ α. Therefore di+1 ≤ di + 1+5α1/3. Combining these two inequalities, we
obtain d d +1+min 9α/d ,5α1/3 .
i+1 i i
≤ {p }
Inductively we prove that d 3j + 7α1/3 + 5α1/3j2/3. For j 2, the inequality is
j
≤ ≤
clear. Now suppose by induction that d 3j +7α1/3 +5α1/3j2/3. If d 2α1/3, we reach
j j
≤ ≤
the desired inequality using the inequality d d + 1 + 5α1/3. Otherwise, we know
j+1 j
≤
that d d + 1 + 9α/d = f(d ) and to find the maximum of the function f(d )
j+1 j j j j
≤ p
over the domain d [2α1/3,j + 7α1/3 + 5α1/3j2/3], we should check f’s critical points,
j
∈
including the endpoints of the domain interval and where f’s derivative is zero. We reach
three values for d : 2α1/3, j+7α1/3+5α1/3j2/3, and 9α 1/3. Because the third value is not
j (cid:0) 4 (cid:1)
in the domain, we just need to check the first two values. The first value is also checked, so
just the second value remains. For the second value, we have d d +1+ 9α/d
j+1 j j
≤ p ≤
j+7α1/3+5α1/3j2/3+1+ 9α j+1+7α1/3+5α1/3j2/3+ 10α j+1+
qj+7α1/3+5α1/3j2/3 ≤ q5α1/3j2/3 ≤